Qwen-Sharp-Chat-Templates: Control Reasoning Depth via Tokens in Qwen Models

This is a Jinja2-based chat template for the Qwen series of LLMs. By inserting specific tags into user messages, you can control the model's reasoning depth and tool-calling behavior.

Including tags such as <|think_xhigh|>, <|think_low|>, and <|think_medium|> in messages allows the template to detect them and inject corresponding instructions into the system prompt. xhigh requires precise verification and alternative review, while low requires concise conclusion derivation. You can also toggle reasoning mode on or off using <|think_on|> and <|think_off|>.

By default, terse mode is enabled to eliminate unnecessary preambles and repetitions, guiding the model to output only the core answer. This behavior can be disabled by setting the terse flag in chat_template_kwargs to false. When calling tools, you can fix the output format by choosing either JSON or XML.

This is suitable for developers who serve Qwen models directly and want to adjust reasoning quality and response speed according to context. Since the model's reasoning depth can be controlled at the prompt level without additional fine-tuning, it is used for A/B testing and adjusting response tone in production environments.